WindowsSandboxServer.exe crashes (unhandled exception in ManagedWindowsVM.Terminate during VM teardown), killing the running guest

[ericstj]

WindowsSandboxServer.exe crashes (unhandled exception in ManagedWindowsVM.Terminate during VM teardown), killing the running guest

Summary

WindowsSandboxServer.exe (the broker/manager process) intermittently terminates
with an unhandled exception thrown from WindowsUdk.Security.Isolation.ManagedWindowsVM.Terminate()
during VM shutdown/teardown. Because the exception is unhandled on a thread‑pool /
event‑callback thread, the entire broker process crashes, which tears down the
running sandbox guest and blocks new launches until things are cleaned up. Reproduces
on long‑running / high‑throughput sandbox sessions.

Two exception variants are observed, both in the same Terminate() call path:

1. System.ObjectDisposedException: Cannot access a disposed object. Object name: 'ObjectReference'.
   — Terminate() is invoked on a VM whose WinRT IObjectReference has already been disposed.
2. System.Runtime.InteropServices.COMException (0x800706BF): The remote procedure call failed and did not execute.
   — the RPC to terminate the VM fails (VM already gone).

Environment

• Windows Sandbox package: MicrosoftWindows.WindowsSandbox_0.8.107.0_x64__cw5n1h2txyewy (v0.8.107.0)
• Faulting module: KERNELBASE.dll 10.0.26100.8875 · Exception code 0xe0434352 (managed) · Fault offset 0xc1ada
• Runtime: .NET 9.0.18 (CoreCLR 9.0.1826.31522)
• Host: Windows 11 (build 10.0.26100), Windows 365 Cloud PC (nested virtualization), 32 logical CPUs
• Workload: long batches of programmatic sandbox runs — each run maps host folders read‑only,
  runs a build/test workload inside the guest, and the guest self‑shuts‑down (shutdown /s)
  on completion. Crashes cluster around guest teardown.

Frequency

3 crashes captured over ~24h of on/off use (Application/​.NET Runtime event 1026 +
Application Error 1000). All three fault at the identical offset in the same
Terminate() path. WER Report IDs: d915b65c-…, 6c131bcd-…, feb03924-….

Stack traces

Variant 1 — ObjectDisposedException (2 of 3 occurrences)

System.AggregateException: One or more errors occurred. (Cannot access a disposed object. Object name: 'ObjectReference'.)
 ---> System.ObjectDisposedException: Cannot access a disposed object. Object name: 'ObjectReference'.
   at WinRT.IObjectReference.<ThrowIfDisposed>g__ThrowObjectDisposedException|47_0()
   at WinRT.IObjectReference.get_ThisPtr()
   at ABI.WindowsUdk.Security.Isolation.IManagedWindowsVMMethods.Terminate(IObjectReference _obj)
   at WindowsUdk.Security.Isolation.ManagedWindowsVM.Terminate()
   at SandboxServer.VM.SandboxVM.Shutdown()
   at System.Threading.ExecutionContext.RunFromThreadPoolDispatchLoop(...)
   --- End of inner exception stack trace ---
   at System.Threading.Tasks.Task.ThrowIfExceptional(Boolean includeTaskCanceledExceptions)
   at System.Threading.Tasks.Task.Wait(Int32 millisecondsTimeout, CancellationToken cancellationToken)
   at System.Threading.Tasks.Task.Wait()
   at SandboxServer.VM.SandboxVMCreator.<>c__DisplayClass4_0.<Create>b__0(Object sender, EventArgs args)
   at System.Diagnostics.Process.RaiseOnExited()
   at System.Diagnostics.Process.CompletionCallback(Object waitHandleContext, Boolean wasSignaled)
   at System.Threading.RegisteredWaitHandle.PerformCallback(Boolean timedOut)
   at System.Threading.ThreadPoolWorkQueue.Dispatch()
   at System.Threading.PortableThreadPool.WorkerThread.WorkerThreadStart()

Variant 2 — COMException 0x800706BF (RPC failed) (1 of 3 occurrences)

System.Runtime.InteropServices.COMException (0x800706BF): The remote procedure call failed and did not execute.
   at WinRT.ExceptionHelpers.<ThrowExceptionForHR>g__Throw|38_0(Int32 hr)
   at ABI.WindowsUdk.Security.Isolation.IManagedWindowsVMMethods.Terminate(IObjectReference _obj)
   at WindowsUdk.Security.Isolation.ManagedWindowsVM.Terminate()
   at SandboxServer.VM.SandboxVM.Shutdown()
   at SandboxServer.VM.SandboxVM.Dispose()
   at SandboxServer.VM.SandboxManager.OnVmTerminated(Object sender, EventArgs args)
   at System.Threading.Tasks.Task.<>c.<ThrowAsync>b__128_1(Object state)
   at System.Threading.QueueUserWorkItemCallback.Execute()
   at System.Threading.ThreadPoolWorkQueue.Dispatch()
   at System.Threading.PortableThreadPool.WorkerThread.WorkerThreadStart()

Analysis

Both variants are a VM‑lifecycle teardown race: SandboxVM.Shutdown() calls
ManagedWindowsVM.Terminate() on a VM that is already being disposed / already gone —

• reached via the VM process's Process.Exited handler (SandboxVMCreator.<Create>b__0)
  racing with disposal → ObjectDisposedException, or
• reached via SandboxVM.Dispose() → SandboxManager.OnVmTerminated after the VM RPC
  endpoint is already torn down → COMException 0x800706BF.

In both cases the exception propagates unhandled on a thread‑pool/callback thread and
crashes the whole broker (rather than being caught and treated as "already terminated").

Suggested fix

Treat "the VM is already gone" as a no‑op during teardown: guard/try‑catch around
ManagedWindowsVM.Terminate() (and the IObjectReference access) in SandboxVM.Shutdown()
/ the Process.Exited and OnVmTerminated handlers, swallowing ObjectDisposedException
and COMException (0x800706BF) (RPC_S_CALL_FAILED_DNE) since the VM is already terminated.

Impact / workaround

Any programmatic/long‑running use of Windows Sandbox is exposed: a mid‑workload crash
kills the guest and loses in‑flight work. Workaround used here is to make the workload
resumable so a crash only costs in‑flight items — but the broker crash itself is the
underlying bug.